Improved Object Detection Method Utilizing YOLOv7-Tiny for Unmanned Aerial Vehicle Photographic Imagery

Linhua Zhang, Ning Xiong, Xinghao Pan, Xiaodong Yue, Peng Wu, Caiping Guo

TL;DR

This paper introduces PDWT-YOLO, an enhanced object detection model based on YOLOv7-tiny, specifically designed for unmanned aerial vehicle (UAV) imagery. By incorporating a dedicated small-object detection layer, a decoupled detection head, and a novel WIoU loss function, PDWT-YOLO significantly improves detection accuracy for objects of all sizes, especially small ones, while maintaining efficient inference speeds.

What Problem Does This Paper Address?

Object detection from UAVs presents a unique set of challenges. These include the presence of numerous small objects, complex backgrounds, varying object scales, and diverse lighting and weather conditions. Traditional object detection models, even efficient ones like YOLOv7-tiny, often struggle to balance the need for high accuracy with the real-time processing demands of edge computing, particularly when dealing with these small and varied objects. This paper aims to bridge this gap by developing a more robust and accurate object detection system for UAV applications.

Methods

The authors propose the PDWT-YOLO algorithm, which builds upon the YOLOv7-tiny architecture with three key innovations:

  • Small-Object Detection Layer (P2): A new detection layer operating on a higher-resolution feature map (160x160) is introduced. This layer is designed to capture finer details of small objects, improving their detection performance and overall recall.
  • Decoupled Detection Head: The original YOLOv7-tiny detection head (IDetect), which combines classification and regression tasks, is replaced with a decoupled head. This separation allows for independent optimization of classification and localization, reducing interference between the two tasks and leading to improved accuracy and faster convergence.
  • Wise Intersection over Union (WIoU) Loss Function: The standard Complete Intersection over Union (CIoU) loss is replaced with WIoU. WIoU is designed to better handle the balance between high- and low-quality bounding box regression samples and emphasizes anchors of average quality, leading to improved regression accuracy and faster network convergence.

Datasets

The primary dataset used for training and evaluation is the VisDrone-2019 dataset. This dataset is notable for its diverse collection of UAV images captured under various scenarios, weather conditions, and lighting. The paper also mentions using the HIT-UAV dataset for an extended experiment to verify the model's generalizability.

Key Insights and Ideas

  • The "small object problem" is a critical bottleneck: The paper strongly emphasizes that detecting small objects is a major hurdle for UAV-based object detection, and their proposed P2 layer directly targets this.
  • Separating tasks improves performance: Decoupling the classification and regression heads, a common theme in recent object detection research, proves effective here in reducing conflicts and boosting accuracy.
  • Loss functions matter for fine-tuning: The choice of loss function significantly impacts regression accuracy and convergence speed. WIoU's ability to dynamically adjust focus based on anchor quality is a key enabler for better performance.
  • Trade-offs are manageable: The PDWT-YOLO model demonstrates that significant accuracy improvements can be achieved without drastically increasing model complexity or inference time, making it suitable for edge devices.

Key Results

The PDWT-YOLO model achieved significant improvements on the VisDrone-2019 dataset:

  • [email protected]:0.95 increased by 5% compared to the baseline YOLOv7-tiny.
  • [email protected] increased by 6.7% compared to the baseline YOLOv7-tiny.
  • The model demonstrated improved performance across various categories, with notable gains in "pedestrian" (11%) and "van" (8%) categories.
  • Qualitative results (Figure 6) show PDWT-YOLO successfully detecting more small objects in challenging far-field views, with improved detection confidence.
  • Ablation studies confirmed that each of the three proposed improvements (decoupled head, P2 layer, WIoU) individually contributed to performance gains, with their combined effect yielding the best results.
  • The model maintained acceptable running speeds, making it practical for real-world applications.

Limitations and Future Work

  • APL reduction: While the overall performance improved, the paper notes a slight decrease in Average Precision for Large objects (APL) in some ablation experiments. The authors suggest this is a trade-off for better small-object detection and overall feature integration.
  • Generalizability: While tested on HIT-UAV, further validation on a wider array of diverse UAV datasets could strengthen the claims of generalizability.
  • Computational Cost: Although improved over some larger models, the paper doesn't explicitly detail the inference time on specific edge hardware, which would be crucial for deployment.

The paper implicitly suggests future work could involve further optimizing the balance between small and large object detection, exploring more advanced feature fusion techniques, and deploying the model on actual UAV hardware to assess real-world performance.

Summary updated

Abstract

In unmanned aerial vehicle photographs, object detection algorithms encounter challenges in enhancing both speed and accuracy for objects of different sizes, primarily due to complex backgrounds and small objects. This study introduces the PDWT-YOLO algorithm, based on the YOLOv7-tiny model, to improve the effectiveness of object detection across all sizes. The proposed method enhances the detection of small objects by incorporating a dedicated small-object detection layer, while reducing the conflict between classification and regression tasks through the replacement of the YOLOv7-tiny model’s detection head (IDetect) with a decoupled head. Moreover, network convergence is accelerated, and regression accuracy is improved by replacing the Complete Intersection over Union (CIoU) loss function with a Wise Intersection over Union (WIoU) focusing mechanism in the loss function. To assess the proposed model’s effectiveness, it was trained and tested on the VisDrone-2019 dataset comprising images captured by various drones across diverse scenarios, weather conditions, and lighting conditions. The experiments show that [email protected]:0.95 and [email protected] increased by 5% and 6.7%, respectively, with acceptable running speed compared with the original YOLOv7-tiny model. Furthermore, this method shows improvement over other datasets, confirming that PDWT-YOLO is effective for multiscale object detection.
BibTeX
@inproceedings{zhang2023improved,
  author = {Linhua Zhang and Ning Xiong and Xinghao Pan and Xiaodong Yue and Peng Wu and Caiping Guo},
  title = {{Improved Object Detection Method Utilizing YOLOv7-Tiny for Unmanned Aerial Vehicle Photographic Imagery}},
  booktitle = {Algorithms},
  year = {2023},
  url = {https://openreview.net/forum?id=0tLtVdfrWA},
}
Figure